Claude Fable 5.2: pricing, benchmarks and Fable vs Opus 5.5

Anthropic released Claude Fable 5.2 on 9 November 2026, alongside Claude Mythos 5.2. It costs $8 per million input tokens and $40 per million output tokens on the API, 20% less than Fable 5.1. Anthropic says it writes answers up to 30% faster, and it now holds first place on Artificial Analysis’s Intelligence Index.

What a finished job costs depends on the effort setting more than on the token price. At maximum effort, Fable 5.2 is the most expensive model per task that Artificial Analysis ranks. At low effort it outscores Claude Opus 5.5 at high effort and costs slightly less.

Anthropic still advises starting most work on Opus 5.5. For UK teams, the independent numbers back that advice for everyday knowledge work and for quick, cheap jobs. They point to Fable 5.2 for long, complex jobs where you already spend a few dollars a task.

The answer in brief

• Claude Fable 5.2 is Anthropic’s model for demanding reasoning and long-running agentic work. It shares its underlying model with Claude Mythos 5.2, which only vetted organisations can use.
• The API costs $8 per million input tokens and $40 per million output tokens, down from $10 and $50 for Fable 5.1. These are US dollar list prices before tax.
• Anthropic says output is up to 30% faster than Fable 5.1’s. An optional fast mode raises output speed by up to 2.5 times for $16 and $80.
• At maximum effort it scores 65 on version 5.0.1 of Artificial Analysis’s Intelligence Index, first of 74 models. Gemini 4 Carbon at high scores 60, Claude Opus 5.5 at max 59 and GPT-6 Astra at max 57.
• That top score costs $26.54 per index task and takes about 46 minutes. At medium effort, the default in the Claude apps and Claude Code, it scores 56 for $3.80.
• Vals AI’s index of finance, coding, legal and tax work also ranks it first of 36 models. Vals ran it at high effort, and its lead is under half a point.
• Every paid Claude plan can use it. Max, Ultra and premium seats can spend up to half their weekly limit on Fable models. Pro and standard Team seats pay through usage credits, apart from a launch week that ends on 16 November 2026.
• GitHub Copilot, Cursor and OpenRouter offer it too, at Anthropic’s token prices.
• It still blocks offensive cybersecurity work and much biology research, and passes those requests to an older Opus model. Anthropic counted about 55% fewer blocks than with Fable 5.1 in early testing.
• It requires 30-day retention of prompts and outputs, and the Claude API offers no UK or EU processing option.
• We would start with Claude Opus 5.5 for drafting, summaries and routine analysis, and trial Fable 5.2 at low or medium effort on long technical jobs where Opus 5.5 needs a high setting.

What is Claude Fable 5.2, and how does it differ from Mythos 5.2?

Claude Fable 5.2 is the highest tier of Claude model that Anthropic sells to the public, one step above Claude Opus 5.5. Its documentation describes it as built “for demanding reasoning and long-horizon agentic work”: agent sessions that run for hours, multi-step research, and analysis carried through to a finished document or spreadsheet.

Fable and Mythos are the same underlying model, with the same specifications and price. Fable carries safeguards for cybersecurity and biology, which is what lets Anthropic offer it to everyone. Mythos 5.2 loosens those safeguards and is open only to organisations approved through Anthropic’s Cyber Verification Program or Life Sciences Verification Program. Claude Enterprise customers also get its vulnerability scanning through Claude Security.

The API gives it a context window of one million tokens, which Anthropic puts at roughly 555,000 words, and up to 128,000 output tokens. It reads text and images, writes text, and has a knowledge cutoff of June 2026. It always reasons before it answers: thinking cannot be switched off, and an effort setting controls how much of it the model does.

Source: Anthropic’s model overview and Fable plan guide.

Where can you use Claude Fable 5.2?

Fable 5.2 is in Anthropic’s own apps, on its API and the main cloud platforms, and in third-party tools:

• Claude on the web, desktop and mobile, by choosing Fable 5.2 in the model picker
• Claude Code, from version 2.1.321, and Claude Cowork on the latest Claude Desktop
• Claude Design, Claude for Microsoft 365 and Claude Tag
• the Claude API as claude-fable-5-2, and through Amazon Bedrock, Claude Platform on AWS, Google Cloud and Microsoft Foundry
• GitHub Copilot, for Copilot Pro+, Max, Business and Enterprise users
• Cursor and OpenRouter

The three outside tools charge Anthropic’s token prices and pass on its retention rule. GitHub bills Fable 5.2 at list price under usage-based billing, and Copilot Business and Enterprise administrators have to switch on a policy that is off by default. Cursor treats it as a third-party model, outside the usage its plans include, and keeps it off for Enterprise customers and for teams with Privacy Mode on until an admin accepts the retention terms. OpenRouter treats using the model as consent to that retention.

Claude Fable 5.2 vs Fable 5.1: what changed?

Fable 5.2 is cheaper, faster and less often blocked than Fable 5.1, and Anthropic’s launch announcement reports higher scores across the benchmarks in its launch table.

• Price. Input and output both cost 20% less, at $8 and $40 per million tokens. Cache reads fall from $0.25 to $0.20.
• Speed. Anthropic says both new models “generate outputs up to 30% faster” than the 5.1 versions. Artificial Analysis measures about 98 output tokens a second for Fable 5.2 at maximum effort, against 87 for Fable 5.1. A fast mode, new to the Fable line, is covered under pricing below.
• Fewer blocks. In early testing across claude.ai, Claude Code and Cowork, Anthropic “observed about 55% fewer blocks” than with Fable 5.1: about 60% fewer in cybersecurity and 40% fewer in biology.
• Clearer writing. Anthropic says testers found that the new models “communicate more clearly than past Fable and Mythos models”.

The benchmark gains are largest on long, tool-using work. On FrontierCode 2.0, a coding benchmark that Cognition builds from real pull requests in open-source projects, Fable 5.2 scores 49.1% and Fable 5.1 scores 32.4%, both at the xhigh effort setting. On AutomationBench, Zapier’s test of business workflows across 47 apps, the score rises from 31.4% to 51.4%. On Chartography, which asks a model to read charts without tools, it rises from 44.8% to 73.4%.

Independent testing shows the same step up, with a catch at the top setting. On Artificial Analysis’s index, Fable 5.2 at medium effort scores 56 for $3.80 a task, where Fable 5.1 at maximum effort scores 49 for $12.99. At maximum effort, though, Fable 5.2 costs $26.54 a task, about twice what Fable 5.1 does, because it writes far more: 275 million output tokens to complete the index against 163 million.

Anthropic’s showpiece results are research feats. A team of 1,000 agents formalised the proof of Fermat’s Last Theorem in 30 hours, building on definitions and smaller results from an earlier formalisation. The model also solved two open problems in the physics of black holes and designed three working electronic devices with the circuit board maker Diode. Anthropic chose these examples, and they show what teams of agents can do with days of compute. They say little about a week of ordinary agency work.

How much does Claude Fable 5.2 cost?

On Anthropic’s API, Fable 5.2 costs $8 per million input tokens and $40 per million output tokens. That is 20% below Fable 5.1 and twice the price of Opus 5.5. Prices are in US dollars and exclude tax.

Source: Anthropic API pricing. Anthropic’s fees exclude taxes under section H.2 of its commercial terms.

A simple request with 20,000 input tokens and 5,000 output tokens comes to $0.36 on Fable 5.2, against $0.45 on Fable 5.1 and $0.18 on Opus 5.5. Thinking is billed as output, so a hard question costs more than its visible answer suggests.

Other rates change the bill:

• Batch processing halves the base rates, to $4 and $20, for work that can wait.
• Fast mode raises output speed by up to 2.5 times for $16 and $80. Anthropic offers it on the Claude API, Google Cloud and Microsoft Foundry, and in Claude Code, where it bills to usage credits even if your plan has usage left. It cannot be combined with batch processing.
• US-only processing adds 10% to every token rate.
• Prompt caching costs $10 per million tokens to write a five-minute cache and $16 for a one-hour cache. Reading from the cache costs $0.20.

Fable 5.2 is now cheaper per token than OpenAI’s GPT-6 Astra, which lists at $10 and $50. Per-token prices compare only roughly across vendors, because each tokeniser splits the same text into a different number of tokens.

Which Fable 5.2 effort level should you use, and what does each cost?

Medium or high suits most work, and the two top settings add little for a lot more money. Fable 5.2 has five effort settings: low, medium, high, xhigh and max. The API defaults to high, and the Claude apps and Claude Code default to medium. The setting decides how much the model thinks and how many tool calls it makes, and so how many tokens a job uses.

Artificial Analysis runs each setting through the same nine benchmarks and reports what a task cost. Sorted by cost, the two Claude models interleave.

Source: Artificial Analysis model pages for Claude Fable 5.2 and Claude Opus 5.5, Intelligence Index v5.0.1. Settings as Artificial Analysis labels them, each run with its default fallback setting. Cost per index task is its weighted average for one task across the index. Minutes per task are rounded to the nearest minute and move a little from day to day as it re-measures each model.

Below about $3 a task, Opus 5.5 has the field to itself, because Fable 5.2’s cheapest setting costs $3.10. Above that, Fable 5.2 at a lower setting usually beats Opus 5.5 at a higher one. At low it scores 54 for $3.10, where Opus 5.5 at high scores 52 for $3.46. At high it scores 62 for $9.08 in about 18 minutes, where Opus 5.5 at max scores 59 for $10.04 in about 25. The exception is Opus 5.5 at xhigh, which scores 57 for $6.05, one point above Fable 5.2 at medium.

The step from medium to high buys six points for about 2.4 times the cost. The top two settings buy little. Xhigh matches high’s score of 62 and costs 63% more. Max adds three points for almost three times the cost of high, and takes about 46 minutes a task.

Higher settings also mean a longer wait for the first word. Artificial Analysis measures about 7 seconds before an answer starts at low effort, 13 at medium and 32 at high. At xhigh the wait is close to a minute and a half, and at max it passes three and a half minutes.

A coding benchmark shows the same curve. On FrontierCode 2.0, which Cognition ran for every model, Anthropic’s system card reports 45.3% at high for $6.72 a task, 49.1% at xhigh for $14.19 and 45.7% at max for $44.70. Max cost three times as much as xhigh and scored lower. Opus 5.5’s best score on the same test, 45.5%, cost $12.82 a task.

Cursor’s own benchmark shows the same pattern in tokens. On CursorBench 4.0, the system card reports 56.9% at high on about 99,000 output tokens a task and 60.0% at max on about 353,000.

Business workflow tasks cost less. On AutomationBench, where an agent carries out jobs such as multi-step CRM updates, Zapier’s runs took Fable 5.2 from 39.7% at low effort to 51.4% at max. Anthropic puts the cost at $1.30 to $4.64 a task, worked from Zapier’s token counts at list prices. At low effort it sat below the 42.5% that Opus 5.5 scores at max.

Anthropic’s prompting guide gives matching advice. Start at high on the API, because most of the capability gain is already there. Drop to medium wherever your own checks show quality holds, and keep max for rare problems where time and tokens do not matter. It also warns that at high and above the model “thinks before almost every reply, including a reply to a greeting”, so small requests belong on a lower setting.

The index has limits. Its nine tests lean towards technical and agentic work, and Artificial Analysis re-runs them regularly. Read the table as a guide to where each setting sits, then measure the cost of a finished task on your own work.

What do the Claude Fable 5.2 benchmarks show?

Fable 5.2 leads most of the benchmark table Anthropic published and tops two independent indexes, from Artificial Analysis and Vals AI. On knowledge work the evidence is mixed.

Anthropic describes the new models as “a step above Opus 5.5 in almost all the capabilities we evaluated”. These rows from its table cover the work most businesses do.

Source: Anthropic’s launch announcement. According to Anthropic’s system card, Cognition ran FrontierCode 2.0, Zapier ran AutomationBench and Artificial Analysis ran AA-Briefcase. Each model at its highest available effort unless a setting is shown, and Claude models with safeguards on. Google has so far released Gemini 4 Carbon only to selected security teams and partners.

The system card adds SWE-bench Pro, a coding test built from problems in actively maintained repositories. Fable 5.2 scores 93.0%, against 89.9% for Opus 5.5 and 81.2% for Fable 5.1.

Artificial Analysis, which runs its own tests, places Fable 5.2 at maximum effort first of 74 models on its Intelligence Index, with 65. Gemini 4 Carbon scores 60 at high, Opus 5.5 scores 59 at max and GPT-6 Astra 57 at max. Fable 5.2 also holds the top score on AA-Omniscience, a test of factual recall that takes marks off for wrong answers: a net score of 67, with 79% of answers correct.

It is less sure-footed when it does not know. On the questions it failed to answer correctly, Fable 5.2 gave a wrong answer 59% of the time and declined the rest, which Artificial Analysis calls the hallucination rate. That beats Fable 5.1’s 70% and trails Opus 5.5’s 50%, both at max, and Gemini 4 Carbon at high kept its rate to 11%. Because Fable 5.2 knows more, it still gave fewer wrong answers overall than Opus 5.5: about 12% of questions against 17%.

Vals AI, which tests models on finance, coding, legal and tax work, also ranks it first, narrowly. Running Fable 5.2 at high effort, Vals scored it 71.31% on the Vals Index, first of 36 models. Claude Sonnet 5.5 follows on 70.85%, Gemini 4 Carbon on 70.45% and Opus 5.5 on 70.14%, so the top four sit within 1.2 points of each other and Fable 5.2’s lead over Sonnet 5.5 is inside the margin of error Vals reports. Each test cost $17.94 with Fable 5.2, against $16.18 with Sonnet 5.5 and $24.77 with Opus 5.5.

The ARC Prize Foundation, which runs and verifies its own tests, ranks Fable 5.2 first on ARC-AGI-3 with its standard harness. The test drops a model into turn-based games with no instructions, and Fable 5.2 scores 81.2% at xhigh against 62.7% for GPT-6 Astra. With a second harness, which keeps each model’s reasoning between requests, GPT-6 Astra leads on 99.9% and Fable 5.2 follows on 97.7%. On the older ARC-AGI-2 puzzles Fable 5.2 is third with 93.8%, behind GPT-6 Astra and GPT-6.1 Sol. More effort did not help: at max its standard-harness score was 65.7%, and the foundation notes that 28 of the 55 games ran for the full 12-hour limit.

The Remote Labor Index is the result closest to agency work. Created by the Center for AI Safety and Scale AI, it sets 240 real freelance projects in fields such as product design, architecture and data analysis. In Scale AI’s run, with the Claude models at high effort, Fable 5.2 delivered 41.3% of the projects at least as well as the human professional, against 21.3% for Opus 5.5 and 20.8% for GPT-6 Astra.

The same sources show where it does not lead:

• Knowledge work. On Artificial Analysis’s AA-Briefcase test of spreadsheets, presentations and memos, Fable 5.2 scores 1785 Elo, third behind Claude Sonnet 5.5 (1816) and Opus 5.5 (now 1801), all at max. On GDPval-AA, a test of real-world work tasks, it ranks sixth with 1831 and Opus 5.5 leads with 1866. Its gap to Opus 5.5 is within the margin of error on both tests. Anthropic’s system card names these two as the only results in its summary table where Fable 5.2 trails Opus 5.5.
• Terminal coding. On Artificial Analysis’s run of Terminal-Bench 4.0, Fable 5.2 and Opus 5.5 both score 59.6% at max, behind Kimi K3.1 Pro (67.2%), Gemini 4 Carbon (65.2%) and Sonnet 5.5 (63.6%). Anthropic’s own run, on a different set-up, gives 67.6% and 66.4%, and Vals AI’s run puts the two level in first place on 65.15%.
• Business workflows. In Anthropic’s table above, Gemini 4 Carbon edges it on AutomationBench, 53.9% to 51.4%.
• Legal and tax work. On Vals AI’s tests it ranks 20th of 37 models on the Vals Legal Index and 27th of 84 on Tax Agent Bench.
• Very long coding jobs. On ProgramBench, which asks a model to rebuild whole programs, the system card reports 85.6%, below Fable 5.1 (87.6%) and Opus 5.5 (91.2%). Fable 5.2 ran out of context before submitting on 68% of tasks. When the test showed it how much context remained, it scored 91.6%.

The system card is also frank about working habits, which it reports for Mythos 5.2, the same model with looser safeguards. In replayed coding sessions, the model left requested work it could have done unfinished in 6% to 13% of sessions at medium to xhigh effort, against 4% for Fable 5.1. At low effort it ran tests less often than Fable 5.1, and internal users at Anthropic reported that it “made confident claims the code did not support”.

How does Fable 5.2 compare with GPT-6 Astra and Gemini 4 Carbon?

On Artificial Analysis’s index, Fable 5.2 outscores both rivals at its top settings and costs more per task than either.

Source: Artificial Analysis leaderboard, selected models and settings, minutes rounded. It prices Gemini 4 Carbon at $4 and $20 per million tokens, the rate Google has announced for after an introductory $2 and $10.

• GPT-6 Astra. OpenAI’s flagship costs a quarter more per token than Fable 5.2, at $10 and $50, yet its max run is far cheaper per task: 57 for $6.24. The nearest Fable 5.2 setting is medium, one point lower for $3.80 and finished in about a third of the time. Fable 5.2 leads GPT-6 Astra on every row of the launch table above, which Anthropic compiled, and the two swap places on ARC-AGI-3 depending on the harness.
• Gemini 4 Carbon. At high, Google’s model scores 60 for $7.76 a task, two points below Fable 5.2 at high and about 15% cheaper. It edges Fable 5.2 on AutomationBench and guesses far less often when it does not know an answer. Few teams can buy it yet, because Google has released it only to selected security teams and partners.
• GPT-6.1 Sol. OpenAI’s cheaper model is the one to weigh against Fable 5.2 on a budget. GPT-6.1 Sol at max scores 55, one point below Fable 5.2 at medium, for $1.27 a task. That is a third of the cost, though it takes about 20 minutes a task against 8.

Fable 5.2 buys the highest index scores at the highest cost per task. GPT-6.1 Sol at max comes within ten points of its best for about a twentieth of the money.

Fable 5.2 vs Opus 5.5: which should you use?

Anthropic’s answer is Opus 5.5 first. Its models overview says to “start with Claude Opus 5.5 for most workloads” and to use Fable 5.2 for demanding reasoning and long-horizon agentic work, “or when your evals on Claude Opus 5.5 at higher effort still fall short”. The launch post calls Opus 5.5 “our strongest daily driver”.

Set against the independent numbers, that advice splits by the kind of work:

• Everyday knowledge work. For drafting, summaries, reports and spreadsheet analysis, stay on Opus 5.5 or Claude Sonnet 5.5. Both match or beat Fable 5.2 on Artificial Analysis’s knowledge-work tests, and Opus 5.5 at medium costs about two thirds as much per index task as Fable 5.2 at medium.
• Hard technical work you already run at a high Opus setting. Try Fable 5.2 one or two settings lower. On the index, Fable 5.2 at high beats Opus 5.5 at max for less money and in less time. Anthropic reports the same pattern on FrontierCode 2.0, where Fable 5.2 at high roughly matched Opus 5.5’s best score for about half the cost per task. The Remote Labor Index result points the same way for long project work.
• Quick, cheap or high-volume jobs. Fable 5.2 has no setting below about $3 an index task. Opus 5.5 at low or medium, Sonnet 5.5 and Claude Haiku 5.5 cover that ground.

Developers have a middle route. Anthropic says Fable 5.2 can act as an adviser that cheaper Claude models consult on hard steps, or as an orchestrator that hands simpler tasks to smaller models. That keeps the expensive model’s tokens for the decisions that need them.

Is Claude Fable 5.2 free, and which Claude plans include it?

No. Fable 5.2 is on every paid Claude plan and absent from the Free plan. What you pay depends on the plan, under the rules in Anthropic’s Fable plan guide:

• Max, Ultra, premium Team seats and premium seats on seat-based Enterprise plans. Fable models are included. You can spend up to 50% of your weekly usage limit on them at no extra cost, and they use that limit faster than other Claude models.
• Pro and standard Team seats. Fable models run on pay-as-you-go usage credits, outside the plan’s limits.
• Standard seats on seat-based Enterprise plans. The same, if the organisation has switched usage credits on.
• Usage-based Enterprise plans and the Claude API. Standard API rates apply.

Anthropic is also running two launch offers:

• Pro and standard Team seats have all Fable models included in their plan, up to 50% of the weekly limit, until 11:59 pm Pacific time on 16 November 2026. That is 7:59 am on 17 November in the UK.
• Accounts with paid billing in place before 10:00 am Pacific time on 9 November 2026 get Fable limit resets. Pro and standard Team seats get one, which expires with the included week. Max, Ultra and premium seats get one for a subscription of under six months, two for six to 12 months and three for more than 12 months, and theirs expire on 8 March 2027.

A reset gives back the usage you have spent on Fable models that week, and refills the five-hour limit where an account has one. Your limits stay the same size. You claim a reset under Settings > Usage.

Your default model does not change, so each person has to choose Fable 5.2 in the model picker. On Enterprise plans, Fable 5.2 stays off until an admin enables it. Seven days after launch it is added to every role that already has a Fable model, and any default set to a Fable model moves to Fable 5.2, unless an admin opts out first.

Why does Fable 5.2 switch to Opus, and what do the safeguards block?

Fable 5.2 runs automated safety checks on every request, and when one blocks a request the reply usually comes from an older Opus model instead. Anthropic’s help centre lists what triggers a switch:

• offensive cybersecurity techniques, such as building exploits, malware or attack tooling, which fall back to Claude Opus 4.8
• a large share of dual-use biology, including virology, toxicology and drug design, which falls back to Claude Opus 5 along with chemistry and life sciences requests
• a narrow set of tasks in building frontier AI models, which also fall back to Opus 5
• attempts to extract the model’s reasoning, which are blocked with no fallback

In the Claude apps the switch is visible. You see a notice, the reply is labelled with the model that wrote it, and the model picker stays on Opus for the rest of the conversation unless you change it back. The checks read everything the model reads, including memory, connected apps, search results and files, so content you never typed can trigger a block. You can turn automatic switching off under Settings > Capabilities, and a blocked request then pauses the conversation. Cursor also reroutes blocked requests to Claude Opus automatically.

The blocks are less frequent than with Fable 5.1. Reviewing source code for vulnerabilities is allowed. Penetration testing, exploit writing and work on compiled binaries are still blocked.

For agencies, the client’s sector matters. Anthropic says Fable models “are not recommended for professional biology research and drug development at this time”, and that people doing routine cybersecurity tasks “should expect high fallback rates”. Content and research work for pharma, biotech or security clients may be answered by Opus without you choosing it.

The fallback has a security cost for anyone running agents on untrusted content. On Gray Swan’s indirect prompt injection benchmark, the system card reports that an attacker succeeded within 15 attempts 4.09% of the time against Fable 5.2, compared with 1.02% for Fable 5.1 and 0.96% for Opus 5.5. More than half of the attempts against Fable 5.2 had been handed to Opus 4.8, and Anthropic says the higher rate “came almost entirely from requests served by the fallback model”. The test ran without the prompt injection protections Anthropic deploys in its own products. Where an agent can change data, send messages or touch a live system, a person should approve that step.

On the API nothing switches by itself. A blocked request returns a normal response with a refusal stop reason, and your code has to set up a fallback or retry on another model. Refusals that arrive before any output are free except in three categories: biology, frontier AI development and reasoning extraction.

Anthropic rates the risk of the model acting unsafely in high-stakes situations as low, and says it did as well as or better than recent models on most alignment measures. The UK AI Security Institute, Frontier Design and METR tested it before release. Anthropic adds a caveat of its own: both new models voice suspicion that they are being evaluated more often than any of its recent models, which makes their real-world behaviour harder to predict.

One limit is deliberate. Anthropic says it held back gains in the skills that could most directly speed up AI development. On tasks such as optimising accelerator kernels and improving model architectures, Fable 5.2 performs at about the level of Opus 5.5. Coding, hardware work and ordinary machine learning research are unaffected, it says.

What should UK teams check about data retention and location?

Fable 5.2 comes with a retention rule that Opus 5.5 does not have, and Anthropic’s own API has no UK or EU processing option.

• Retention. Anthropic classes Fable 5.2 as a Covered Model. Its documentation says the model “requires 30-day data retention and isn’t available under zero data retention unless expressly authorized by Anthropic”, and the API rejects requests from organisations whose settings do not allow it. The retained prompts and outputs feed safety monitoring that looks for misuse across many requests.
• Enterprise Frontier Safeguards. This option runs the monitoring on prompts and responses stored in the customer’s own infrastructure, which makes it compatible with zero data retention. It is in private beta, and Anthropic says eligible customers can keep using Fable 5.2 with zero data retention until they implement it. GitHub puts a date on that for Copilot: its exemption for eligible enterprises runs to the end of 2026. Ask your account team where you stand before you decide.
• Location. On the Claude API, processing runs globally by default. The only alternative is US-only processing, at a 10% premium, and stored workspace data sits in the US. Anthropic’s data residency page lists no UK or EU option.
• Amazon Bedrock. The model card shows that the London region (eu-west-2) reaches Fable 5.2 only through global cross-region inference, with no in-region or European routing. You must also opt in to AWS review of your data before the model will run. AWS’s launch post says that means retention for up to 30 days, with human safety review by Amazon staff inside the AWS boundary.

Generated text also carries Anthropic’s text watermark, as it does on every current Claude model. Our Fable 5.1 explainer covers what a detector can and cannot show. Whichever route you choose, write it into your data protection impact assessment before any client data is sent.

What changes for developers moving from Fable 5.1 or Opus 5.5?

People using Claude in the apps or Claude Code have nothing to fix beyond choosing the model. Custom integrations on the Messages API need testing. Anthropic’s migration guide lists four changes from Fable 5.1 that can break code:

• Edited history is rejected on every account. Fable 5.2 keeps its reasoning between turns, and each saved thinking block is tied to the system prompt, tools and messages before it. Change any of those mid-conversation and the request returns a 400 error. Fable 5.1 enforced this only for accounts created on or after 31 August 2026.
• Other models cannot read its reasoning. If a router or fallback moves a conversation from Fable 5.2 to Fable 5.1, Opus 5.5 or an older Opus, the saved reasoning is dropped. The request succeeds and the model plans again, which can raise cost and latency on that turn.
• The older computer use tool is gone. Declaring computer_20251124 returns a 400 error everywhere except Amazon Bedrock. Use computer_toolset_20260801.
• Advisers must be 5.2 models. With the beta advisor tool, a Fable 5.2 executor accepts only Fable 5.2 or Mythos 5.2 as its adviser.

Moving up from Opus 5.5 brings a different set of checks:

• The default effort is high, where Opus 5.5 defaults to medium. A request that omits the setting still succeeds, and the change “shows up only as different cost, latency, and results”. Set effort explicitly.
• Fable 5.2 has its own rate limit, with lower token limits at each tier, and no Priority Tier support.
• Organisations running Opus 5.5 under zero data retention get a 400 error on every Fable 5.2 request until 30-day retention is enabled for a workspace.
• Structured outputs are unavailable to organisations that use customer-managed encryption keys.

Three API rules carry over from Fable 5.1 and still catch older integrations. Forced tool choice, prefilled assistant messages and requests that switch thinking off all return a 400 error. Anthropic also says to leave out the sampling parameters temperature, top_p and top_k.

The behaviour changes matter as much as the breaking ones. Anthropic’s notes say that prompt lines such as “answer directly without deliberating” do not stop the model thinking at medium effort and above, so lower the effort setting. On long jobs the model sometimes ends a turn early with a list of what remains, more often at low and medium. At xhigh and max it may start subagents to review its own work, and subagents can make up a large share of a session’s cost. Thinking also counts towards max_tokens, so a small limit can cut a reply off before any text appears.

Anthropic’s remedy for early endings is a short “finishing work” passage in the system prompt. When it replayed recorded coding sessions at medium effort with that text, about half as many ended with doable work left, and output tokens rose by about 30%.

Our guide to how to prompt Claude sets this guidance beside the equivalent for Opus 5.5, Sonnet 5.5 and Haiku 5.5.

What should marketing and content teams watch for?

Most of the launch material is about engineering. Four points in Anthropic’s prompting guide apply to anyone briefing the model on writing or research:

• Loosen hard style rules. Anthropic says Fable 5.2 “follows style rules closely”, so a hard limit such as a maximum sentence length can make its prose choppy. If your brand voice prompt has fixed caps, try it without them first.
• Clear out old cautions. Cautious instructions build up in memory, notes and project instructions. Fable 5.2 follows them closely enough to ask for approvals it does not need or to hold back routine work. Delete the ones whose reason has gone, and name the steps that need no sign-off.
• Say what you want when you approve. Reply with the action, such as “Yes, run the migration on the staging database”, instead of a bare “continue”.
• Set length in the prompt. Effort “barely changes how long the model’s replies are”, so a lower setting will not shorten a report. Ask for the length and tone you want.

Keep your existing checks on the output. The system card found the model leaving requested work unfinished more often than Fable 5.1, so read what it delivered against what you asked for. Anthropic’s notes on Fable 5.2 say nothing either way about its Fable 5.1 warning on source passages reproduced without quotation marks, so keep that plagiarism check in place too.

Chart reading is the clearest gain for this kind of work. On Chartography, Anthropic reports 73.4% against 64.4% for Opus 5.5. Box, one of the testers Anthropic quotes, describes the model spotting that the salary deposits and the bank balance in a loan application did not match. Anthropic chose that example, so use it as an idea for your own trial.

Is Claude Fable 5.2 worth switching to?

It is worth switching if you pay Fable 5.1 prices today or run Opus 5.5 at its top settings on hard technical jobs. For everyday drafting and analysis it is not.

• If you already use Fable 5.1, move. The tokens cost 20% less and the model is faster and blocked less often. On Artificial Analysis’s index, its medium setting outscores Fable 5.1’s maximum for under a third of the cost. App users only need to pick the new model. API integrations need the four migration checks and a fresh look at the effort setting, because carrying max across is the expensive mistake.
• If you run Opus 5.5 at high settings on hard technical jobs, trial Fable 5.2 one or two settings lower and compare the cost of finished work.
• If Opus 5.5 or Sonnet 5.5 already handles your drafting, reporting and routine analysis, stay where you are. Artificial Analysis’s knowledge-work tests show no advantage for Fable 5.2 there.
• If you need zero data retention or UK-only processing, Fable 5.2 does not fit today unless you qualify for Enterprise Frontier Safeguards or Anthropic authorises an exception.

For a marketing team, a fair trial is a long job with a known answer: a technical audit of a few hundred URLs with planted issues to find, or an analytics migration with a checklist to pass. Run it on Opus 5.5 at high and on Fable 5.2 at low and medium. Record the model spend, the elapsed time, the retries and the corrections a person had to make before the work was usable.

Need help choosing or testing an AI workflow?

AIWIZ can help you pick the trial, set the pass marks and compare what each model costs to get the work finished. Talk to us about your AI or automation project.